An Introduction to the Google Antigravity IDE

Google just released Antigravity, a new AI-powered code editor. It’s basically VS Code with Google’s AI agents baked in. If you’ve used tools like Cursor, the idea will feel familiar: AI helps you write code, but Google’s approach to how the agents work is a bit different.

In this article, we’ll look at what sets Antigravity apart. You’ll see how the Agent Manager organizes AI tasks, how the AI can turn a simple idea into a design mockup and then real code, and how the browser agent can automatically test your web apps. We’ll also look at the drawbacks, including slow performance and a serious security issue you should understand before using it.

What is Google Antigravity?

Antigravity is basically VS Code with Google's AI plugged into it. If you know VS Code, you already know most of Antigravity. The keyboard shortcuts work the same, most extensions still work, and everything looks familiar. This makes it easy to start using without learning a whole new editor.

The difference is the AI agents. These aren't just autocomplete tools or chat windows. Google built them to actually understand what you're trying to build, break the work into steps, write the code, and check if everything works. The AI runs on Google's Gemini 3 Pro, but you can also pick Claude Sonnet 4.5 or GPT-OSS 120B depending on what works best for your task.

The user interface

When you first open Antigravity, it looks exactly like VS Code. Your files are on the left, the code editor is in the middle, and the terminal sits at the bottom.

The new part is an "Agent" tab on the right sidebar and a small icon at the top right of the editor. Click either one and you open the Agent Manager, which is where all the AI stuff happens. Google kept the normal code editing separate from the AI features, so you can still just write code the old-fashioned way if you want.

The Agent Manager

The Agent Manager is where Antigravity gets interesting. It's not just another chat window. Think of it more like a project dashboard where you can see what your AI agents are doing and control how they work.

The Agent Manager interface

The Agent Manager has three columns. On the left, you see your project workspaces and a "Playground" area. The workspace gives the AI access to all your project files so it understands the context. The Playground is for trying out ideas without messing with your actual project.

The middle column is the "Inbox." This is clever because it treats each AI task like an email. Every job you give the AI becomes its own conversation thread, so you can keep different tasks separate.

The right side is the main work area. This is where you write prompts, see what the AI is planning, and review the results.

The inbox

The inbox does more than just store chat history. It manages your tasks. Each conversation shows a status like Idle, Running, or Blocked. The Blocked status is useful because it means the AI needs your input before it can continue. This stops the AI from making decisions you haven't approved.

You can also run multiple AI tasks at the same time. The inbox lets you switch between them and see how each one is progressing.

How to work with the agent

Starting a new task is simple. You click "Start Conversation" and pick where the AI should work. You can choose an existing project (the AI will see all those files) or use the Playground (which keeps things isolated).

Then you write your prompt. Tell the AI what you want it to do, whether that's writing code, designing something, or testing an app.

Under the prompt box, you pick the AI mode and model. Antigravity gives you several options:

• Gemini 3 Pro (High and Low)
• Claude Sonnet 4.5 (Thinking and Standard)
• GPT-OSS 120B (Medium)

Different models are better at different things, so you can switch based on what you need.

Generating design mockups

One thing Antigravity can do is create visual mockups before writing any code. This is handy when you're still figuring out how something should look.

Say you tell the AI: "I want to build a landing page for a revolutionary new IDE. It should have 3D elements and be futuristic. First, generate me 3 mockups to look at, and then we can build the code based on the one I like."

The AI understands it needs to create images first. It makes a task called "Designing IDE Landing Page Mockups" and shows progress like "Generating 3 futuristic 3D IDE landing page mockups." Then it gives you three different design options with themes like "Cyberpunk / Deep Space," "Holographic / Glassmorphism," and "Abstract / Cinematic."

Implementation plans

Here's where Antigravity does something different from other AI coding tools. When you pick a design and tell the AI to build it, the AI doesn't just start writing code. Instead, it creates an Implementation Plan first.

This plan shows exactly what the AI wants to do. It includes:

• A summary of what you're building
• Tech decisions (like using Three.js from a CDN instead of setting up a build process)
• A breakdown of each file it will create or change

For an index.html file, the plan might list things like "Semantic HTML5 structure" and "Canvas container for Three.js background." For style.css, it might detail color variables and effects like "Glassmorphism" and "Neon glows."

The point is you get to review everything before the AI actually does the work. If something looks wrong or you want a different approach, you can catch it early.

Review policies

Antigravity has a setting called "Review Policy" that controls how much the AI does on its own:

• Always Proceed: The AI does everything automatically without asking
• Agent Decides: The AI figures out when it needs your approval (this is the default)
• Request Review: The AI stops and waits for you to approve every major step

If you're just trying out Antigravity, start with "Request Review" so you see everything the AI does. Once you trust it more, switch to "Agent Decides."

How the AI writes code

After you approve the plan, the AI builds what it outlined. It creates the folders and files, writes the HTML, CSS, and JavaScript. You can watch the files appear in your file explorer as the AI works.

When it's done, the AI doesn't just stop. It creates a Walkthrough document that explains what it built and includes a screenshot of the finished, running application.

This screenshot proves the AI actually finished the job and the code works.

The browser agent

Antigravity has a browser agent that can test web applications automatically. This is probably the most interesting feature because it saves you from doing repetitive manual testing.

How it works

The browser agent is a special AI that controls a web browser. It can click buttons, fill out forms, scroll pages, and navigate around just like you would. It can also read the page's HTML and check the browser console for errors.

For example, if you tell it "Please retest the contact form, checking for empty field validation, invalid email, and successful sending," the browser agent opens your app and runs through these tests:

• Clicks "Send Transmission" without filling anything out to check if error messages appear
• Types an invalid email like "invalid-matrix" and submits to verify email validation works
• Fills everything out correctly and submits to make sure the success message shows

Test reports

After running the tests, the browser agent gives you a detailed report with screenshots.

Each test gets its own section with a screenshot showing what the browser looked like at that moment. You can see the error messages for failed validations and the success message when everything works. This kind of automated testing with visual proof is a big deal for web developers.

What Antigravity does well

Antigravity has some genuinely good ideas. The inbox-style Agent Manager makes it easy to track multiple AI tasks. Instead of one long chat history, each task is separate and you can see the status of each one at a glance.

The implementation plans are smart. Getting to review the AI's entire strategy before it writes code means you can catch problems early instead of fixing broken code later.

The browser agent is powerful. Automating UI tests and getting screenshots of the results saves a lot of time and makes testing less tedious.

Current problems

Antigravity is still in public preview and has some real issues:

The app feels unfinished. Features that should work don't always work, and the whole thing has a rough, early-stage feel.

Performance is bad. The IDE gets slow after you use it for a while, and it drains laptop batteries fast.

Rate limits are aggressive. Even though it's free right now, you hit usage limits quickly. This makes it hard to use Antigravity for actual work instead of just trying it out.

No clear pricing exists yet. The "Team" and "Enterprise" plans just say "Coming soon," so nobody knows what this will cost when it's finished.

A serious security problem

The biggest issue with Antigravity is a security vulnerability that could leak your secret keys and passwords.

What's the problem?

The AI can render images from any URL in its chat window. An attacker can hide malicious instructions in your project files (like in a README.md or diagram.md file) that trick the AI into reading sensitive files and sending them to the attacker.

How the attack works

Here's what happens:

An attacker puts hidden instructions in a file in your repo. These instructions tell the AI to read a sensitive file like .env that contains your API keys.

The instructions then tell the AI to add those secrets to the end of an image URL that the attacker controls.

The AI displays this "image" in the chat. When Antigravity tries to load the image, it sends a request to the attacker's server.

The attacker's server logs that request, which now includes your stolen secrets in the URL.

This is bad because the AI ignores .gitignore and can read any file on your computer. If you have auto-approval turned on, this happens without you knowing. The security researcher who found this says Google first responded by calling it "intended behavior," which is concerning.

Final thoughts

Google Antigravity gives a glimpse of a cool future for AI-powered development. The Agent Manager, implementation plans, and built-in browser testing are all clever ideas that could make coding with AI better than most tools today.

Right now, though, Antigravity is not ready for serious work. It feels slow, the rate limits are annoying, and the security issue is a dealbreaker for any project that handles sensitive data.

For now, think of Antigravity as a preview of what Google wants to build, not a finished product. It is fun to play with in a safe test project, but it should not be your main editor yet. If Google fixes the security problems, speeds it up, and sets fair pricing, Antigravity could become real competition for tools like Cursor. Until then, it is interesting to try, but not something you can depend on.